How to define an ICP that an AI can prospect against

An ideal customer profile describes the type of company that gets the most value from your product and gives you the most value back: it buys, stays, expands and is reasonable to serve. It describes companies, not people. The people inside those companies are covered by buyer roles, which sit on top of the ICP. Mixing the two is the most common source of confusion, because "marketing managers at tech companies" says something about a role and almost nothing about which companies to target.

An ICP is also not your total addressable market. The market is everyone who could in theory buy. The ICP is the narrow slice where you win most often and fastest. A tight ICP feels uncomfortable because it leaves out companies you could sell to, and that discomfort is the point: outbound works when every message is relevant, and relevance comes from narrowness.

An experienced rep fills gaps with judgment. Told to target "growing SaaS companies", they quietly skip a two-person side project and a public company with ten thousand staff. An AI prospecting against the same sentence has no such instinct unless you give it criteria. It will find companies that match the words, and the words allow almost anything.

So the ICP for AI prospecting needs three properties. Each criterion must be observable in data the system can read: a website, a company register, a job post, a news item, your own CRM. Each criterion must have a clear boundary, such as a range for headcount or a list of countries. And the profile must say what to exclude, because exclusions remove most of the noise.

If you have customers, the best ICP is hiding in them. Pull a list of your strongest accounts, the ones that bought quickly, use the product well, renew and refer others. Ignore the logos you are proud of if they took a year to close and need constant attention. For each strong account, write down a handful of facts.

• Industry and what they sell, in plain words.
• Headcount band and, if you know it, revenue band.
• Country and region where the buying team sits.
• Business model, for example subscription software, services, manufacturing, marketplace.
• Who signed and who used the product day to day.
• What was happening at the company when they bought.

Patterns appear quickly, often after ten or fifteen accounts. If you have no customers yet, use the same template for the companies you believe are the best fit and mark the profile as a hypothesis to be tested, not a fact.

Firmographics are the company level facts: industry, size, location, model and technology. They form the outer filter of your ICP. Write them as ranges and lists rather than adjectives.

• Industry: name the specific segment, such as "freight forwarding and third party logistics", not "logistics".
• Size: give a headcount band, such as 50 to 500 employees, because headcount is visible far more often than revenue.
• Geography: list countries or regions, and note language if your outreach is in one language only.
• Business model: say whether you sell to companies that sell software, services, physical goods or something else.
• Technology or setup: include it only when it truly matters, for example "uses a cloud CRM" or "runs an online store".
• Stage: if relevant, describe it in observable terms, such as "has a sales team" or "has raised outside funding", rather than "fast growing".

Avoid criteria that only exist in your head. "Innovative", "sophisticated" and "serious about growth" cannot be checked, so a machine will either ignore them or guess.

Once you know which companies to target, decide whom to write to. In most business purchases, several people are involved: someone who feels the problem every day, someone who owns the budget, and sometimes someone who can block the purchase, such as security, legal or IT. Your outbound should usually start with the person who owns the problem and has enough authority to start a conversation.

• Primary role: the title family most likely to reply, for example VP Operations or Head of Revenue Operations.
• Secondary roles: who else to contact if the primary role does not exist at a smaller company, often the founder or a general manager.
• Seniority range: which levels to include, because very junior contacts cannot move a purchase and very senior ones rarely read cold email at large companies.
• Roles to avoid: titles that look relevant but rarely buy, such as interns or roles in an unrelated department with a similar name.

Write titles as families rather than exact strings. "Operations leadership: VP, Director or Head of Operations" covers the variations real companies use.

Firmographics tell you who could buy. Triggers tell you who might buy now. A trigger is an observable event that makes your product more relevant this quarter than last. Triggers are what give an AI something real to say in the first line, so they deserve as much care as the firmographics.

• Hiring for a role your product supports or replaces, visible in job posts.
• A new leader in the buying role, who often reviews tools in their first months.
• Expansion into a new region, product line or customer segment.
• Funding, acquisition or a public announcement of growth plans.
• A change on their website, such as new pricing, a new product page or a new market.
• A regulatory or industry change that creates new work for their team.

Pick the two or three triggers that most often preceded purchases among your best customers. A company in your ICP with an active trigger should be at the top of your list. A company with no trigger can still be contacted, but the message has to work harder.

Exclusions are the fastest way to improve a prospect list. They are also the part most teams forget. List the companies and situations you never want in a sequence.

• Existing customers and open opportunities, matched against your CRM.
• Competitors and their close partners.
• Companies outside your legal or operational reach, such as countries you cannot serve.
• Segments that match the firmographics but rarely buy, such as public sector bodies if your sales process cannot handle tenders.
• Companies that have unsubscribed or asked not to be contacted, through one shared suppression list.
• Personal or consumer contacts, because cold B2B outreach should reach business contacts only.

After all of the above, compress the profile into a single sentence. One line forces choices, and it gives the AI a clear brief while letting you add detail underneath. A useful structure is: we sell [offer] to [industry] companies in [geography] with [size band], where the [primary role] is dealing with [problem], especially when [trigger], excluding [exclusions].

Filled in, it reads like this: we sell freight audit software to US third party logistics companies with 50 to 500 employees, where the VP Operations is dealing with invoice disputes, especially when they are hiring billing staff or expanding to new lanes, excluding current customers and carriers. A person can read that sentence and picture ten companies. So can a prospecting system.

Do not send to a list just because it matches the words. Ask the system, or a colleague, to produce ten to twenty matching companies, then review them by hand. For each company, ask three questions: would we be happy to have this customer, does the trigger really apply, and is the named role the right person to contact. Mark each one as a good fit, a weak fit or a mistake.

Mistakes point to missing exclusions or vague criteria. Weak fits point to a band that is too wide or a trigger that is too loose. Rewrite the one-line ICP after each round. Two or three rounds usually produce a profile where almost every company on the list is one you would want to talk to.

It is tempting to start broad so the pipeline looks full. Start narrow instead. A small, precise segment lets you write messages that sound like you know the reader's world, and those messages earn replies. Once a narrow segment is producing conversations, widen one dimension at a time: a neighbouring industry, the next size band, another country. Keep each variation as a separate segment so you can see which one works, rather than mixing them into one large list where results blur together.

Your ICP changes as your product and market change. Review it every quarter against new customers and lost deals. Look at which segments booked meetings, which replied "not now" and which went silent. Move budget and attention towards the segments that convert, and write down why you dropped the others so nobody quietly adds them back.

• Offer in one sentence, in the buyer's words.
• Industry segment, as specific as possible.
• Headcount band and geography.
• Business model and any setup that matters.
• Primary role, secondary roles and roles to avoid.
• Two or three buying triggers that can be observed.
• Exclusions, including customers, competitors and suppressed contacts.
• The one-line ICP sentence that combines all of the above.